April 24, 2025
\[ \newcommand\hbb{{\hat{\boldsymbol \beta}}} \newcommand\bb{{\boldsymbol \beta}} \newcommand\expn{{\frac{1}{N} \sum \limits_{i = 1}^N}} \newcommand\sumk{\sum \limits_{k = 1}^K} \newcommand\argminb{\underset{\bb}{\text{argmin }}} \newcommand\argmaxb{\underset{\bb}{\text{argmax }}} \newcommand\gtheta{\mathbf g(\boldsymbol \theta)} \newcommand\htheta{\mathbf H(\boldsymbol \theta)} \]
If you want your poster shown off at the event tomorrow, get me your poster ASAP.
I really think that having an opportunity to show off will really put you in a place where you can realize just how much stuff you know
If you want to present your poster tomorrow, come to the 2nd floor lobby in PAIS starting at 11:30
You don’t have to stand by your poster the whole time
Not a requirement, but a good way to get some clout
Final deliverable: Paper or website for your project
Due May 7th
Need to see that you actually did something
Should serve as a base for a paper or project that you can continue later
Can’t be late. Last day to grade and enter grades for graduating seniors
What I care about the most.
I’ll hold regular office hours up until the 7th.
Modern AI methods are accessible
Most have no idea how they work and why they do what they do.
A new ethical code is needed for highly accessible generative AI
Fake information has gotten so much more realistic looking in the last two years
It’s not only harm that can be caused at a macro level.
Individual’s can easily be harmed:
Personalized cyberbullying
Deepfakes
Catfishing
Sextortion
Is there a general set of rules for how we should leverage/gate modern generative AI?
A big question.
Do philosophical frameworks of ethics help us here?
An ethical framework centered on outcomes
Make choices that do the greatest good for the greatest number of people
A minimax problem:
Maximize happiness
Minimize harm
Open-sourcing powerful generative AI models like Stable Diffusion
Strengths:
Practical and outcome focused
Encourages consideration of broad societal impacts
Clear criterion (maximize welfare)
Weaknesses:
Requires quantification
Rewards the majority, penalizes the minority
Can justify unethical means to a “positive” end
An ethical framework focused on duties, rights, and universal ethical rules
Contraints placed on the means, not the ends.
Ethical actions respect universal moral duties
Honesty
Fairness
Consent
Open-sourcing powerful generative AI models like Stable Diffusion
Strengths:
Clarity in ethical duties
Protects individual rights
Promote consistency
Weaknesses:
Sometimes the end justifies the means!
No conclusions possible when duties conflict
An ethical framework focusing on moral character and virtues
What would a morally virtuous person do?
What do you think are some moral pillars that we should consider in an AI model?
In our craw, we know the difference between right and wrong
Decisions must be made based upon our ability to reason through the moral implications.
Do our best to meet a set of moral goods.
Problem: Questions lead to more questions!
Fairness
AI should produce things that promote equity
The world we want, not the world we live in
Treat individuals and groups without discrimination
AI amplifies existing societal biases if unchecked
Unfair outcomes erode trust in AI systems
How can we address problems of fairness in AI?
Do you think that this is enough?
How can we know if our training set will lead to bias?
How can we correct it?
Truthfulness
AI systems should produce content and information that accurately represents reality and transparently discloses its artificial origin
This one is harder:
Turthfulness is a combination of intent and algorithmic design
More complexity = more truthfulness
More complexity = more ability to create untruths in a convincing way
Proposals?
Consent
AI should respect individual autonomy by obtaining explicit permission before collecting, using, or generating representations of their data or identity.
Struggles
How do we collect consent?
Is Instagram fair game? You did consent.
Do IP laws apply to AI?
Github Copilot:
Trained on all code in Github repos
All code is there
Even proprietary code
But, Github has some sense of ownership
Thoughts?
AI Art:
How are we able to replicate a copyrighted style?
How can we prevent this?
Does it stifle the capability of AI art?
Is that a bad thing?
Open-sourcing powerful generative AI models like Stable Diffusion
We don’t have good answers for this!
This is up to you.
A new wave of technically gifted scholars that understand the tools
How they work
Why they work
No quick answers. Just more and more thought needed.
We’ve covered a lot this semester!
Machine learning theory:
Complexity theory (VC dimensionality)
Bias variance tradeoffs
Bayesian statistics
Regularization
Universal Approximation Theory
Convex optimization and stochastic minimization
Basic Neural Networks:
Multilayer Perceptrons
Deep Neural Networks
Activations for Nonlinearities
Generalization Methods
NNs for Images:
CNNs
Residual Networks and Batch Normalization
Modern CNN architectures
Bounding box detection
Semantic segmentation and UNets
NNs for Sequences:
RNNs
LSTMs
Seq2Seq
Attention
Self-Attention and Transformers
BERT and GPT
Deep Generative Methods:
Autoregressive Models
Variational Autoencoders
Generative Adversarial Networks
Diffusion Models
The pace of this class was incredibly ambitious and you all kept up with the materials admirably!
Above all else, I hope that you feel like you learned something new:
NN theory above application
Ins and Outs of image analysis
Transformers and Stable Diffusion are just clever combinations of simpler things!
As a second go at this class, there was a mixture of good and bad (from my perspective)
Good:
Lecture order flowed pretty well
Y’all were quite engaged with the material
We covered very modern topics
A broad overview of what we can do with NNs
Good projects!
As a second go at this class, there was a mixture of good and bad (from my perspective)
Bad:
The homeworks got slowed down by resource issues/ability to put together meaningful examples that could run in finite time that weren’t too simple
Homeworks should be more like guided projects than anything. But, that takes too much time. Gotta figure it out.
I’ve learned that I don’t actually like PyTorch Lightning as much as I thought I would and am going to switch to base PyTorch for next year.
It’s been a learning process…
On that note, please fill out the course eval
It matters a lot for my career
Your comments good and bad (though not personal ones about the fact that I only have two sweatshirts and two pairs of jeans) are really valuable for my development as a teacher and the overall development of this class
I have really enjoyed teaching this class
I thought I knew a lot before the semester started, but I learned sooooo much prepping this material!
This is my favorite class I’ve ever designed from the ground up
I’m looking forward to doing this again!
To the graduating seniors:
Congrats!
This is my fourth year here, so I met some of y’all in my first year teaching
You are all so impressive and I know you’re going to do amazing things with your lives!
To those not graduating:
Please don’t be a stranger
I’ll still be around next year, so please stop by and let me know how things are going
I have genuinely enjoyed getting to know each and every one of you
You are all incredibly bright and pretty good programmers
I am really optimistic about the next wave of data scientists :)